Facial thermal ROI tracking for temperature measurement

Tracking a face in a thermal video and measuring its temperature are separate tasks. A tracker locates anatomical regions; a radiometric data stream supplies the measurements inside them. For a research or engineering prototype, you need both, plus rules for rejecting unreliable frames. The workflow below defines five facial regions, follows their motion, and produces regional temperature histories without confusing a display image with measurement data.

Define the measurement and its limits

1. Write a measurement specification before choosing a tracking method. Define the left and right supraorbital regions, left and right periorbital regions, and nose region using anatomical boundaries that someone else can reproduce.

Use the subject's left and right, not the displayed image's left and right. Record whether the preview is mirrored. Supraorbital regions sit above the eye sockets; periorbital regions surround the eyes. These labels alone do not define usable masks, especially where the regions meet.

Document which skin areas belong to each region. Decide whether eyelids, eyebrows, and the inner eye corners are included. Define the nose region as a particular surface, such as the bridge or tip, rather than allowing its coverage to change between recordings.

Facial surface temperature is not a direct measurement of core body temperature. Air movement, sweating, recent activity, environmental conditions, and local circulation can change the measured surface temperature. This workflow does not establish a fever-screening or diagnostic method.

For recordings involving people, obtain appropriate consent and follow applicable privacy and research requirements. Store identifiers separately from exported measurements whenever possible.

Confirm that the camera supplies measurement data

2. Verify the complete radiometric acquisition path. You need documented temperature values or raw measurements with the manufacturer's documented conversion procedure. A colored thermal preview is not sufficient.

A palette image maps measurements into display colors. Automatic contrast adjustment can change those colors even when the underlying measurement stays constant. Do not estimate temperature from displayed color or brightness unless the instrument explicitly documents that measurement method.

Before selecting among infrared cameras, confirm these requirements:

• Access to the measurement array for every acquired frame.
• Documented units, scaling, invalid-value codes, and calibration conditions.
• A way to associate each array with its timestamp and displayed image.
• Suitable optics and working distance for separating the facial regions.
• Access to measurement settings and correction-status information where available.

The camera finder helps organize camera selection, but a category listing does not prove software compatibility or radiometric access. Treat the LS-IRC-0044 as a catalog candidate requiring those checks.

The LS-IRC-0026 is another catalog candidate to discuss against the same requirements. Its listing alone does not establish suitability for facial measurements.

For detector sensitivity and optical tradeoffs, see thermal camera NETD and lens choice for process monitoring. Sensitivity, spatial detail, and absolute temperature accuracy are different specifications.

Capture a repeatable baseline sequence

3. Stabilize acquisition before adding automatic tracking. Start with a fixed camera, a stationary subject, and manually marked regions. This separates measurement problems from tracking problems.

Follow the instrument's documented warm-up and calibration procedures. Record the relevant measurement settings, including emissivity and reflected apparent temperature where the instrument uses them. Do not adjust these settings merely to make the reported skin temperature match an expected value.

Keep camera position, subject distance, background, and airflow reasonably consistent. Document the environmental conditions and any preparation period used in your protocol. There is no universal settling time that makes every facial measurement valid.

Glasses can obscure the intended skin surface and contribute reflected radiation. Hair, hands, and clothing can also replace skin within a region. Treat these as visibility problems rather than expecting the tracker to measure through them.

Save the original measurement arrays separately from annotated previews. Preserve acquisition timestamps, frame identifiers, and acquisition settings so that you can reproduce the calculation later.

Create five masks with explicit anatomical boundaries

4. Initialize each region on a reference frame. A polygon or binary mask is usually more useful than a rectangle because it can exclude adjacent features and background.

For each region, store its label, reference geometry, and inclusion rules. Keep supraorbital and periorbital masks separate if independent measurements are required. If overlap is intentional, record it because the resulting measurements will share pixels.

A practical mask-building sequence is:

• Identify the face and the landmarks needed to define the regions.
• Construct the region boundaries from those landmarks.
• Restrict the boundaries to visible skin.
• Remove invalid pixels and detected obstructions.
• Optionally shrink the mask to reduce contamination near uncertain edges.

Thermal appearance does not necessarily provide the facial detail expected by a detector trained on visible images. Validate any detector on representative thermal recordings before using its landmarks for measurement.

For an initial prototype, manual initialization followed by short-term tracking is often easier to debug than fully automatic detection. If the subject moves substantially, pause or reacquire rather than allowing a mask to drift onto a different anatomical surface.

Mask size also matters. A region containing only a few pixels can change sharply when its boundary shifts slightly. Record the pixel count for every region and frame, not just the temperature result.

Track motion without changing the measurement target

5. Update each mask using a motion model appropriate to the movement. A translated rectangle may work for small sideways motion, but it cannot describe head rotation, facial expression, or partial occlusion.

A useful pipeline combines landmark updates with short-term tracking. Landmarks restore anatomical alignment; tracking supplies continuity between successful detections. Neither should silently override evidence that the region is no longer visible.

Consider an illustrative stream acquired at 25 frames/s. If a facial feature moves across the image at 200 pixels/s, its expected displacement is:

This is a planning example, not a tracker capability claim. A search area must accommodate the expected movement, while remaining selective enough to avoid nearby similar structures.

Update masks from anatomical geometry where possible, rather than repeatedly selecting the warmest patch. Otherwise, a changing hotspot can move the measurement away from the intended tissue.

Define a tracking state for each region: valid, uncertain, occluded, or lost. Reinitialize after a substantial pose change or failed landmark check. Do not keep reporting a stationary mask after the face has moved away.

Map every mask into the measurement array

6. Express the final masks in the coordinates of the radiometric frame. The preview, detector input, and measurement array may have different dimensions, crops, or orientations.

If a detector receives an image resized directly from the thermal frame, account for both horizontal and vertical scale. For example, a preview width of 320 pixels and a measurement-array width of 640 pixels give a horizontal scale factor of:

That calculation is valid only when there is no additional crop, padding, mirror, or rotation. With letterboxing, subtract the padding offset before scaling. With a crop, restore the crop origin as well.

Map polygon vertices into measurement coordinates and rasterize the final mask there. Avoid resizing temperature arrays merely to match the display: interpolation changes values and can alter extrema.

If landmarks come from a separate visible camera, a simple scale factor is generally insufficient. Viewpoint differences, facial depth, calibration, and synchronization all matter. The guide to thermal and RGB image fusion without losing temperature data explains why display alignment must remain separate from radiometric measurement.

Check the mapping by drawing mask outlines over a preview generated from the same measurement frame. Inspect alignment at the image center and edges, not just at one position.

Calculate regional statistics from valid pixels

7. Intersect the anatomical mask with a measurement-validity mask. Exclude documented invalid values, saturated measurements, and obstructed areas before calculating regional statistics.

Let T denote the temperature array and M the final binary mask. The regional mean is:

If N is zero, return a missing result and a reason code. Do not substitute zero temperature or repeat the previous value.

For a worked example, suppose the intended mask contains 240 pixels. Obstructions or invalid measurements remove 40 pixels:

Here, 7,000 °C represents the arithmetic sum of the included Celsius readings, not a physical surface temperature. Whether the remaining coverage is acceptable depends on your protocol and which part of the region was removed.

Choose statistics according to the measurement question:

Store several statistics if useful, but nominate the primary one before analyzing the recording. More decimal places do not establish better measurement accuracy.

Reject unreliable frames and export an auditable history

8. Apply quality gates before smoothing or plotting results. Good-looking curves can conceal a tracker that has lost the anatomy.

For each region, evaluate landmark confidence, visible coverage, pose, boundary position, and valid pixel count. Set acceptance thresholds using labeled recordings from the intended setup. A confidence threshold copied from another detector or application has no guaranteed meaning here.

Export one record per region per frame, including:

• Frame identifier and acquisition timestamp.
• Region label and tracking state.
• Selected temperature statistics and units.
• Intended pixel count, valid pixel count, and valid fraction.
• Geometry or a reference to the stored mask.
• Relevant measurement settings and rejection reason.

Missing values should remain visible as gaps. Interpolation may be acceptable for a separate presentation layer, but it should not replace the original measurement record.

Temporal averaging can reduce short-term fluctuations while delaying real changes. For example, a trailing window of 10 frames at 25 frames/s spans 0.4 s. Label filtered results and retain the unfiltered series for investigation.

Do not mix rejected frames into an average. Also distinguish an acquisition interruption from a tracking failure: both create missing measurements, but they require different fixes.

Verify the pipeline before longer recordings

Validation should test the geometry and temperature path independently. A stable temperature curve alone does not prove that a mask stayed on the intended skin.

Use a short evaluation set containing still poses, translations, head turns, blinking, and deliberate obstructions. Manually annotate representative frames and compare the tracked boundaries with those references.

As a sensitivity check, shift a mask by an illustrative 2 pixels in each direction and recalculate its statistic. A large change suggests that boundary placement, spatial sampling, or adjacent tissue is dominating the result. The acceptable change must come from your measurement requirements.

Check frame timing with dropped-frame and delayed-frame cases. Confirm that measurement arrays are paired with the correct geometry, especially when detection runs asynchronously.

Where quantitative temperature accuracy matters, verify the instrument against an appropriate traceable reference using its documented procedure. A reference target can check the measurement chain, but it cannot turn skin temperature into core temperature or validate the anatomical tracker.

The usual tracking approach is wrong when the required region is hidden, spatially unresolved, or outside the calibrated measurement conditions. Reposition the setup, revise the protocol, or report no measurement instead of forcing a numerical result.

Checklist before selecting parts or recording data

• Measurement access: Confirm documented radiometric output through the intended acquisition interface.
• Optical coverage: Demonstrate that each anatomical region contains enough usable pixels at the planned distance.
• Coordinate mapping: Test crop, scale, orientation, and any separate-camera registration.
• Tracking recovery: Define what happens after occlusion, rapid motion, or landmark failure.
• Measurement consistency: Keep anatomical definitions and selected statistics unchanged across comparable recordings.
• Data integrity: Retain original arrays, timestamps, masks, settings, and rejection reasons.
• Human-data handling: Establish consent, access controls, and retention requirements before collection.

Start with manual masks and a stationary recording, then add motion tracking only after the measurement path passes these checks. Use the camera configuration step to outline the hardware requirements, or contact the applications team with your intended working distance, region definitions, motion range, and radiometric-access requirements before selecting parts.